1 //===-- PoolAllocate.cpp - Pool Allocation Pass ---------------------------===//
3 // This transform changes programs so that disjoint data structures are
4 // allocated out of different pools of memory, increasing locality and shrinking
7 // This pass requires a DCE & instcombine pass to be run after it for best
10 //===----------------------------------------------------------------------===//
12 #include "llvm/Transforms/IPO/PoolAllocate.h"
13 #include "llvm/Transforms/CloneFunction.h"
14 #include "llvm/Analysis/DataStructure.h"
15 #include "llvm/Analysis/DataStructureGraph.h"
16 #include "llvm/Module.h"
17 #include "llvm/Function.h"
18 #include "llvm/BasicBlock.h"
19 #include "llvm/iMemory.h"
20 #include "llvm/iTerminators.h"
21 #include "llvm/iPHINode.h"
22 #include "llvm/iOther.h"
23 #include "llvm/DerivedTypes.h"
24 #include "llvm/Constants.h"
25 #include "llvm/Target/TargetData.h"
26 #include "llvm/Support/InstVisitor.h"
27 #include "llvm/Argument.h"
28 #include "Support/DepthFirstIterator.h"
29 #include "Support/STLExtras.h"
32 // DEBUG_CREATE_POOLS - Enable this to turn on debug output for the pool
33 // creation phase in the top level function of a transformed data structure.
35 //#define DEBUG_CREATE_POOLS 1
37 // DEBUG_TRANSFORM_PROGRESS - Enable this to get lots of debug output on what
38 // the transformation is doing.
40 //#define DEBUG_TRANSFORM_PROGRESS 1
42 // DEBUG_POOLBASE_LOAD_ELIMINATOR - Turn this on to get statistics about how
43 // many static loads were eliminated from a function...
45 #define DEBUG_POOLBASE_LOAD_ELIMINATOR 1
47 #include "Support/CommandLine.h"
49 Ptr8bits, Ptr16bits, Ptr32bits
52 static cl::Enum<enum PtrSize> ReqPointerSize("ptrsize", 0,
53 "Set pointer size for -poolalloc pass",
54 clEnumValN(Ptr32bits, "32", "Use 32 bit indices for pointers"),
55 clEnumValN(Ptr16bits, "16", "Use 16 bit indices for pointers"),
56 clEnumValN(Ptr8bits , "8", "Use 8 bit indices for pointers"), 0);
58 static cl::Flag DisableRLE("no-pool-load-elim", "Disable pool load elimination after poolalloc pass", cl::Hidden);
60 const Type *POINTERTYPE;
62 // FIXME: This is dependant on the sparc backend layout conventions!!
63 static TargetData TargetData("test");
65 static const Type *getPointerTransformedType(const Type *Ty) {
66 if (PointerType *PT = dyn_cast<PointerType>(Ty)) {
68 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
69 vector<const Type *> NewElTypes;
70 NewElTypes.reserve(STy->getElementTypes().size());
71 for (StructType::ElementTypes::const_iterator
72 I = STy->getElementTypes().begin(),
73 E = STy->getElementTypes().end(); I != E; ++I)
74 NewElTypes.push_back(getPointerTransformedType(*I));
75 return StructType::get(NewElTypes);
76 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
77 return ArrayType::get(getPointerTransformedType(ATy->getElementType()),
78 ATy->getNumElements());
80 assert(Ty->isPrimitiveType() && "Unknown derived type!");
87 DSNode *Node; // The node this pool allocation represents
88 Value *Handle; // LLVM value of the pool in the current context
89 const Type *NewType; // The transformed type of the memory objects
90 const Type *PoolType; // The type of the pool
92 const Type *getOldType() const { return Node->getType(); }
94 PoolInfo() { // Define a default ctor for map::operator[]
95 cerr << "Map subscript used to get element that doesn't exist!\n";
99 PoolInfo(DSNode *N, Value *H, const Type *NT, const Type *PT)
100 : Node(N), Handle(H), NewType(NT), PoolType(PT) {
101 // Handle can be null...
102 assert(N && NT && PT && "Pool info null!");
105 PoolInfo(DSNode *N) : Node(N), Handle(0), NewType(0), PoolType(0) {
106 assert(N && "Invalid pool info!");
108 // The new type of the memory object is the same as the old type, except
109 // that all of the pointer values are replaced with POINTERTYPE values.
110 NewType = getPointerTransformedType(getOldType());
114 // ScalarInfo - Information about an LLVM value that we know points to some
115 // datastructure we are processing.
118 Value *Val; // Scalar value in Current Function
119 PoolInfo Pool; // The pool the scalar points into
121 ScalarInfo(Value *V, const PoolInfo &PI) : Val(V), Pool(PI) {
122 assert(V && "Null value passed to ScalarInfo ctor!");
126 // CallArgInfo - Information on one operand for a call that got expanded.
128 int ArgNo; // Call argument number this corresponds to
129 DSNode *Node; // The graph node for the pool
130 Value *PoolHandle; // The LLVM value that is the pool pointer
132 CallArgInfo(int Arg, DSNode *N, Value *PH)
133 : ArgNo(Arg), Node(N), PoolHandle(PH) {
134 assert(Arg >= -1 && N && PH && "Illegal values to CallArgInfo ctor!");
137 // operator< when sorting, sort by argument number.
138 bool operator<(const CallArgInfo &CAI) const {
139 return ArgNo < CAI.ArgNo;
143 // TransformFunctionInfo - Information about how a function eeds to be
146 struct TransformFunctionInfo {
147 // ArgInfo - Maintain information about the arguments that need to be
148 // processed. Each CallArgInfo corresponds to an argument that needs to
149 // have a pool pointer passed into the transformed function with it.
151 // As a special case, "argument" number -1 corresponds to the return value.
153 vector<CallArgInfo> ArgInfo;
155 // Func - The function to be transformed...
158 // The call instruction that is used to map CallArgInfo PoolHandle values
159 // into the new function values.
163 TransformFunctionInfo() : Func(0), Call(0) {}
165 bool operator<(const TransformFunctionInfo &TFI) const {
166 if (Func < TFI.Func) return true;
167 if (Func > TFI.Func) return false;
168 if (ArgInfo.size() < TFI.ArgInfo.size()) return true;
169 if (ArgInfo.size() > TFI.ArgInfo.size()) return false;
170 return ArgInfo < TFI.ArgInfo;
173 void finalizeConstruction() {
174 // Sort the vector so that the return value is first, followed by the
175 // argument records, in order. Note that this must be a stable sort so
176 // that the entries with the same sorting criteria (ie they are multiple
177 // pool entries for the same argument) are kept in depth first order.
178 stable_sort(ArgInfo.begin(), ArgInfo.end());
181 // addCallInfo - For a specified function call CI, figure out which pool
182 // descriptors need to be passed in as arguments, and which arguments need
183 // to be transformed into indices. If Arg != -1, the specified call
184 // argument is passed in as a pointer to a data structure.
186 void addCallInfo(DataStructure *DS, CallInst *CI, int Arg,
187 DSNode *GraphNode, map<DSNode*, PoolInfo> &PoolDescs);
189 // Make sure that all dependant arguments are added to this transformation
190 // info. For example, if we call foo(null, P) and foo treats it's first and
191 // second arguments as belonging to the same data structure, the we MUST add
192 // entries to know that the null needs to be transformed into an index as
195 void ensureDependantArgumentsIncluded(DataStructure *DS,
196 map<DSNode*, PoolInfo> &PoolDescs);
200 // Define the pass class that we implement...
201 struct PoolAllocate : public Pass {
202 const char *getPassName() const { return "Pool Allocate"; }
205 switch (ReqPointerSize) {
206 case Ptr32bits: POINTERTYPE = Type::UIntTy; break;
207 case Ptr16bits: POINTERTYPE = Type::UShortTy; break;
208 case Ptr8bits: POINTERTYPE = Type::UByteTy; break;
211 CurModule = 0; DS = 0;
212 PoolInit = PoolDestroy = PoolAlloc = PoolFree = 0;
215 // getPoolType - Get the type used by the backend for a pool of a particular
216 // type. This pool record is used to allocate nodes of type NodeType.
218 // Here, PoolTy = { NodeType*, sbyte*, uint }*
220 const StructType *getPoolType(const Type *NodeType) {
221 vector<const Type*> PoolElements;
222 PoolElements.push_back(PointerType::get(NodeType));
223 PoolElements.push_back(PointerType::get(Type::SByteTy));
224 PoolElements.push_back(Type::UIntTy);
225 StructType *Result = StructType::get(PoolElements);
227 // Add a name to the symbol table to correspond to the backend
228 // representation of this pool...
229 assert(CurModule && "No current module!?");
230 string Name = CurModule->getTypeName(NodeType);
231 if (Name.empty()) Name = CurModule->getTypeName(PoolElements[0]);
232 CurModule->addTypeName(Name+"oolbe", Result);
239 // getAnalysisUsage - This function requires data structure information
240 // to be able to see what is pool allocatable.
242 virtual void getAnalysisUsage(AnalysisUsage &AU) const {
243 AU.addRequired(DataStructure::ID);
247 // CurModule - The module being processed.
250 // DS - The data structure graph for the module being processed.
253 // Prototypes that we add to support pool allocation...
254 Function *PoolInit, *PoolDestroy, *PoolAlloc, *PoolAllocArray, *PoolFree;
256 // The map of already transformed functions... note that the keys of this
257 // map do not have meaningful values for 'Call' or the 'PoolHandle' elements
258 // of the ArgInfo elements.
260 map<TransformFunctionInfo, Function*> TransformedFunctions;
262 // getTransformedFunction - Get a transformed function, or return null if
263 // the function specified hasn't been transformed yet.
265 Function *getTransformedFunction(TransformFunctionInfo &TFI) const {
266 map<TransformFunctionInfo, Function*>::const_iterator I =
267 TransformedFunctions.find(TFI);
268 if (I != TransformedFunctions.end()) return I->second;
273 // addPoolPrototypes - Add prototypes for the pool functions to the
274 // specified module and update the Pool* instance variables to point to
277 void addPoolPrototypes(Module *M);
280 // CreatePools - Insert instructions into the function we are processing to
281 // create all of the memory pool objects themselves. This also inserts
282 // destruction code. Add an alloca for each pool that is allocated to the
285 void CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
286 map<DSNode*, PoolInfo> &PoolDescs);
288 // processFunction - Convert a function to use pool allocation where
291 bool processFunction(Function *F);
293 // transformFunctionBody - This transforms the instruction in 'F' to use the
294 // pools specified in PoolDescs when modifying data structure nodes
295 // specified in the PoolDescs map. IPFGraph is the closed data structure
296 // graph for F, of which the PoolDescriptor nodes come from.
298 void transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
299 map<DSNode*, PoolInfo> &PoolDescs);
301 // transformFunction - Transform the specified function the specified way.
302 // It we have already transformed that function that way, don't do anything.
303 // The nodes in the TransformFunctionInfo come out of callers data structure
304 // graph, and the PoolDescs passed in are the caller's.
306 void transformFunction(TransformFunctionInfo &TFI,
307 FunctionDSGraph &CallerIPGraph,
308 map<DSNode*, PoolInfo> &PoolDescs);
313 // isNotPoolableAlloc - This is a predicate that returns true if the specified
314 // allocation node in a data structure graph is eligable for pool allocation.
316 static bool isNotPoolableAlloc(const AllocDSNode *DS) {
317 if (DS->isAllocaNode()) return true; // Do not pool allocate alloca's.
321 // processFunction - Convert a function to use pool allocation where
324 bool PoolAllocate::processFunction(Function *F) {
325 // Get the closed datastructure graph for the current function... if there are
326 // any allocations in this graph that are not escaping, we need to pool
327 // allocate them here!
329 FunctionDSGraph &IPGraph = DS->getClosedDSGraph(F);
331 // Get all of the allocations that do not escape the current function. Since
332 // they are still live (they exist in the graph at all), this means we must
333 // have scalar references to these nodes, but the scalars are never returned.
335 vector<AllocDSNode*> Allocs;
336 IPGraph.getNonEscapingAllocations(Allocs);
338 // Filter out allocations that we cannot handle. Currently, this includes
339 // variable sized array allocations and alloca's (which we do not want to
342 Allocs.erase(remove_if(Allocs.begin(), Allocs.end(), isNotPoolableAlloc),
346 if (Allocs.empty()) return false; // Nothing to do.
348 #ifdef DEBUG_TRANSFORM_PROGRESS
349 cerr << "Transforming Function: " << F->getName() << "\n";
352 // Insert instructions into the function we are processing to create all of
353 // the memory pool objects themselves. This also inserts destruction code.
354 // This fills in the PoolDescs map to associate the alloc node with the
355 // allocation of the memory pool corresponding to it.
357 map<DSNode*, PoolInfo> PoolDescs;
358 CreatePools(F, Allocs, PoolDescs);
360 #ifdef DEBUG_TRANSFORM_PROGRESS
361 cerr << "Transformed Entry Function: \n" << F;
364 // Now we need to figure out what called functions we need to transform, and
365 // how. To do this, we look at all of the scalars, seeing which functions are
366 // either used as a scalar value (so they return a data structure), or are
367 // passed one of our scalar values.
369 transformFunctionBody(F, IPGraph, PoolDescs);
375 //===----------------------------------------------------------------------===//
377 // NewInstructionCreator - This class is used to traverse the function being
378 // modified, changing each instruction visit'ed to use and provide pointer
379 // indexes instead of real pointers. This is what changes the body of a
380 // function to use pool allocation.
382 class NewInstructionCreator : public InstVisitor<NewInstructionCreator> {
383 PoolAllocate &PoolAllocator;
384 vector<ScalarInfo> &Scalars;
385 map<CallInst*, TransformFunctionInfo> &CallMap;
386 map<Value*, Value*> &XFormMap; // Map old pointers to new indexes
389 Instruction *I; // Instruction to update
390 unsigned OpNum; // Operand number to update
391 Value *OldVal; // The old value it had
393 RefToUpdate(Instruction *i, unsigned o, Value *ov)
394 : I(i), OpNum(o), OldVal(ov) {}
396 vector<RefToUpdate> ReferencesToUpdate;
398 const ScalarInfo &getScalarRef(const Value *V) {
399 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
400 if (Scalars[i].Val == V) return Scalars[i];
402 cerr << "Could not find scalar " << V << " in scalar map!\n";
403 assert(0 && "Scalar not found in getScalar!");
408 const ScalarInfo *getScalar(const Value *V) {
409 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
410 if (Scalars[i].Val == V) return &Scalars[i];
414 BasicBlock::iterator ReplaceInstWith(Instruction *I, Instruction *New) {
415 BasicBlock *BB = I->getParent();
416 BasicBlock::iterator RI = find(BB->begin(), BB->end(), I);
417 BB->getInstList().replaceWith(RI, New);
422 LoadInst *createPoolBaseInstruction(Value *PtrVal) {
423 const ScalarInfo &SC = getScalarRef(PtrVal);
424 vector<Value*> Args(3);
425 Args[0] = ConstantUInt::get(Type::UIntTy, 0); // No pointer offset
426 Args[1] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of pool descriptr
427 Args[2] = ConstantUInt::get(Type::UByteTy, 0); // Field #0 of poolalloc val
428 return new LoadInst(SC.Pool.Handle, Args, PtrVal->getName()+".poolbase");
433 NewInstructionCreator(PoolAllocate &PA, vector<ScalarInfo> &S,
434 map<CallInst*, TransformFunctionInfo> &C,
435 map<Value*, Value*> &X)
436 : PoolAllocator(PA), Scalars(S), CallMap(C), XFormMap(X) {}
439 // updateReferences - The NewInstructionCreator is responsible for creating
440 // new instructions to replace the old ones in the function, and then link up
441 // references to values to their new values. For it to do this, however, it
442 // keeps track of information about the value mapping of old values to new
443 // values that need to be patched up. Given this value map and a set of
444 // instruction operands to patch, updateReferences performs the updates.
446 void updateReferences() {
447 for (unsigned i = 0, e = ReferencesToUpdate.size(); i != e; ++i) {
448 RefToUpdate &Ref = ReferencesToUpdate[i];
449 Value *NewVal = XFormMap[Ref.OldVal];
452 if (isa<Constant>(Ref.OldVal) && // Refering to a null ptr?
453 cast<Constant>(Ref.OldVal)->isNullValue()) {
454 // Transform the null pointer into a null index... caching in XFormMap
455 XFormMap[Ref.OldVal] = NewVal = Constant::getNullValue(POINTERTYPE);
456 //} else if (isa<Argument>(Ref.OldVal)) {
458 cerr << "Unknown reference to: " << Ref.OldVal << "\n";
459 assert(XFormMap[Ref.OldVal] &&
460 "Reference to value that was not updated found!");
464 Ref.I->setOperand(Ref.OpNum, NewVal);
466 ReferencesToUpdate.clear();
469 //===--------------------------------------------------------------------===//
470 // Transformation methods:
471 // These methods specify how each type of instruction is transformed by the
472 // NewInstructionCreator instance...
473 //===--------------------------------------------------------------------===//
475 void visitGetElementPtrInst(GetElementPtrInst *I) {
476 assert(0 && "Cannot transform get element ptr instructions yet!");
479 // Replace the load instruction with a new one.
480 void visitLoadInst(LoadInst *I) {
481 Instruction *PoolBase = createPoolBaseInstruction(I->getOperand(0));
483 // Cast our index to be a UIntTy so we can use it to index into the pool...
484 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
485 Type::UIntTy, I->getOperand(0)->getName());
487 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I->getOperand(0)));
489 vector<Value*> Indices(I->idx_begin(), I->idx_end());
490 Instruction *IdxInst =
491 BinaryOperator::create(Instruction::Add, Indices[0], Index,
492 I->getName()+".idx");
493 Indices[0] = IdxInst;
494 Instruction *NewLoad = new LoadInst(PoolBase, Indices, I->getName());
496 // Replace the load instruction with the new load instruction...
497 BasicBlock::iterator II = ReplaceInstWith(I, NewLoad);
499 // Add the pool base calculator instruction before the load...
500 II = NewLoad->getParent()->getInstList().insert(II, PoolBase) + 1;
502 // Add the idx calculator instruction before the load...
503 II = NewLoad->getParent()->getInstList().insert(II, Index) + 1;
505 // Add the cast before the load instruction...
506 NewLoad->getParent()->getInstList().insert(II, IdxInst);
508 // If not yielding a pool allocated pointer, use the new load value as the
509 // value in the program instead of the old load value...
512 I->replaceAllUsesWith(NewLoad);
515 // Replace the store instruction with a new one. In the store instruction,
516 // the value stored could be a pointer type, meaning that the new store may
517 // have to change one or both of it's operands.
519 void visitStoreInst(StoreInst *I) {
520 assert(getScalar(I->getOperand(1)) &&
521 "Store inst found only storing pool allocated pointer. "
524 Value *Val = I->getOperand(0); // The value to store...
525 // Check to see if the value we are storing is a data structure pointer...
526 //if (const ScalarInfo *ValScalar = getScalar(I->getOperand(0)))
527 if (isa<PointerType>(I->getOperand(0)->getType()))
528 Val = Constant::getNullValue(POINTERTYPE); // Yes, store a dummy
530 Instruction *PoolBase = createPoolBaseInstruction(I->getOperand(1));
532 // Cast our index to be a UIntTy so we can use it to index into the pool...
533 CastInst *Index = new CastInst(Constant::getNullValue(POINTERTYPE),
534 Type::UIntTy, I->getOperand(1)->getName());
535 ReferencesToUpdate.push_back(RefToUpdate(Index, 0, I->getOperand(1)));
537 vector<Value*> Indices(I->idx_begin(), I->idx_end());
538 Instruction *IdxInst =
539 BinaryOperator::create(Instruction::Add, Indices[0], Index, "idx");
540 Indices[0] = IdxInst;
542 Instruction *NewStore = new StoreInst(Val, PoolBase, Indices);
544 if (Val != I->getOperand(0)) // Value stored was a pointer?
545 ReferencesToUpdate.push_back(RefToUpdate(NewStore, 0, I->getOperand(0)));
548 // Replace the store instruction with the cast instruction...
549 BasicBlock::iterator II = ReplaceInstWith(I, Index);
551 // Add the pool base calculator instruction before the index...
552 II = Index->getParent()->getInstList().insert(II, PoolBase) + 2;
554 // Add the indexing instruction...
555 II = Index->getParent()->getInstList().insert(II, IdxInst) + 1;
557 // Add the store after the cast instruction...
558 Index->getParent()->getInstList().insert(II, NewStore);
562 // Create call to poolalloc for every malloc instruction
563 void visitMallocInst(MallocInst *I) {
564 const ScalarInfo &SCI = getScalarRef(I);
568 if (!I->isArrayAllocation()) {
569 Args.push_back(SCI.Pool.Handle);
570 Call = new CallInst(PoolAllocator.PoolAlloc, Args, I->getName());
572 Args.push_back(I->getArraySize());
573 Args.push_back(SCI.Pool.Handle);
574 Call = new CallInst(PoolAllocator.PoolAllocArray, Args, I->getName());
577 ReplaceInstWith(I, Call);
580 // Convert a call to poolfree for every free instruction...
581 void visitFreeInst(FreeInst *I) {
582 // Create a new call to poolfree before the free instruction
584 Args.push_back(Constant::getNullValue(POINTERTYPE));
585 Args.push_back(getScalarRef(I->getOperand(0)).Pool.Handle);
586 Instruction *NewCall = new CallInst(PoolAllocator.PoolFree, Args);
587 ReplaceInstWith(I, NewCall);
588 ReferencesToUpdate.push_back(RefToUpdate(NewCall, 1, I->getOperand(0)));
591 // visitCallInst - Create a new call instruction with the extra arguments for
592 // all of the memory pools that the call needs.
594 void visitCallInst(CallInst *I) {
595 TransformFunctionInfo &TI = CallMap[I];
597 // Start with all of the old arguments...
598 vector<Value*> Args(I->op_begin()+1, I->op_end());
600 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i) {
601 // Replace all of the pointer arguments with our new pointer typed values.
602 if (TI.ArgInfo[i].ArgNo != -1)
603 Args[TI.ArgInfo[i].ArgNo] = Constant::getNullValue(POINTERTYPE);
605 // Add all of the pool arguments...
606 Args.push_back(TI.ArgInfo[i].PoolHandle);
609 Function *NF = PoolAllocator.getTransformedFunction(TI);
610 Instruction *NewCall = new CallInst(NF, Args, I->getName());
611 ReplaceInstWith(I, NewCall);
613 // Keep track of the mapping of operands so that we can resolve them to real
615 Value *RetVal = NewCall;
616 for (unsigned i = 0, e = TI.ArgInfo.size(); i != e; ++i)
617 if (TI.ArgInfo[i].ArgNo != -1)
618 ReferencesToUpdate.push_back(RefToUpdate(NewCall, TI.ArgInfo[i].ArgNo+1,
619 I->getOperand(TI.ArgInfo[i].ArgNo+1)));
621 RetVal = 0; // If returning a pointer, don't change retval...
623 // If not returning a pointer, use the new call as the value in the program
624 // instead of the old call...
627 I->replaceAllUsesWith(RetVal);
630 // visitPHINode - Create a new PHI node of POINTERTYPE for all of the old Phi
633 void visitPHINode(PHINode *PN) {
634 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
635 PHINode *NewPhi = new PHINode(POINTERTYPE, PN->getName());
636 for (unsigned i = 0, e = PN->getNumIncomingValues(); i != e; ++i) {
637 NewPhi->addIncoming(DummyVal, PN->getIncomingBlock(i));
638 ReferencesToUpdate.push_back(RefToUpdate(NewPhi, i*2,
639 PN->getIncomingValue(i)));
642 ReplaceInstWith(PN, NewPhi);
645 // visitReturnInst - Replace ret instruction with a new return...
646 void visitReturnInst(ReturnInst *I) {
647 Instruction *Ret = new ReturnInst(Constant::getNullValue(POINTERTYPE));
648 ReplaceInstWith(I, Ret);
649 ReferencesToUpdate.push_back(RefToUpdate(Ret, 0, I->getOperand(0)));
652 // visitSetCondInst - Replace a conditional test instruction with a new one
653 void visitSetCondInst(SetCondInst *SCI) {
654 BinaryOperator *I = (BinaryOperator*)SCI;
655 Value *DummyVal = Constant::getNullValue(POINTERTYPE);
656 BinaryOperator *New = BinaryOperator::create(I->getOpcode(), DummyVal,
657 DummyVal, I->getName());
658 ReplaceInstWith(I, New);
660 ReferencesToUpdate.push_back(RefToUpdate(New, 0, I->getOperand(0)));
661 ReferencesToUpdate.push_back(RefToUpdate(New, 1, I->getOperand(1)));
663 // Make sure branches refer to the new condition...
664 I->replaceAllUsesWith(New);
667 void visitInstruction(Instruction *I) {
668 cerr << "Unknown instruction to FunctionBodyTransformer:\n" << I;
673 // PoolBaseLoadEliminator - Every load and store through a pool allocated
674 // pointer causes a load of the real pool base out of the pool descriptor.
675 // Iterate through the function, doing a local elimination pass of duplicate
676 // loads. This attempts to turn the all too common:
678 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
679 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
680 // %reg109.poolbase23 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
681 // store double %reg207, %root.p* %reg109.poolbase23, uint %reg109, ...
684 // %reg109.poolbase22 = load %root.pool* %root.pool, uint 0, ubyte 0, ubyte 0
685 // %reg207 = load %root.p* %reg109.poolbase22, uint %reg109, ubyte 0, ubyte 0
686 // store double %reg207, %root.p* %reg109.poolbase22, uint %reg109, ...
689 class PoolBaseLoadEliminator : public InstVisitor<PoolBaseLoadEliminator> {
690 // PoolDescValues - Keep track of the values in the current function that are
691 // pool descriptors (loads from which we want to eliminate).
693 vector<Value*> PoolDescValues;
695 // PoolDescMap - As we are analyzing a BB, keep track of which load to use
696 // when referencing a pool descriptor.
698 map<Value*, LoadInst*> PoolDescMap;
700 // These two fields keep track of statistics of how effective we are, if
701 // debugging is enabled.
703 unsigned Eliminated, Remaining;
705 // Compact the pool descriptor map into a list of the pool descriptors in the
706 // current context that we should know about...
708 PoolBaseLoadEliminator(const map<DSNode*, PoolInfo> &PoolDescs) {
709 Eliminated = Remaining = 0;
710 for (map<DSNode*, PoolInfo>::const_iterator I = PoolDescs.begin(),
711 E = PoolDescs.end(); I != E; ++I)
712 PoolDescValues.push_back(I->second.Handle);
714 // Remove duplicates from the list of pool values
715 sort(PoolDescValues.begin(), PoolDescValues.end());
716 PoolDescValues.erase(unique(PoolDescValues.begin(), PoolDescValues.end()),
717 PoolDescValues.end());
720 #ifdef DEBUG_POOLBASE_LOAD_ELIMINATOR
721 void visitFunction(Function *F) {
722 cerr << "Pool Load Elim '" << F->getName() << "'\t";
724 ~PoolBaseLoadEliminator() {
725 unsigned Total = Eliminated+Remaining;
727 cerr << "removed " << Eliminated << "["
728 << Eliminated*100/Total << "%] loads, leaving "
729 << Remaining << ".\n";
733 // Loop over the function, looking for loads to eliminate. Because we are a
734 // local transformation, we reset all of our state when we enter a new basic
737 void visitBasicBlock(BasicBlock *) {
738 PoolDescMap.clear(); // Forget state.
741 // Starting with an empty basic block, we scan it looking for loads of the
742 // pool descriptor. When we find a load, we add it to the PoolDescMap,
743 // indicating that we have a value available to recycle next time we see the
744 // poolbase of this instruction being loaded.
746 void visitLoadInst(LoadInst *LI) {
747 Value *LoadAddr = LI->getPointerOperand();
748 map<Value*, LoadInst*>::iterator VIt = PoolDescMap.find(LoadAddr);
749 if (VIt != PoolDescMap.end()) { // We already have a value for this load?
750 LI->replaceAllUsesWith(VIt->second); // Make the current load dead
753 // This load might not be a load of a pool pointer, check to see if it is
754 if (LI->getNumOperands() == 4 && // load pool, uint 0, ubyte 0, ubyte 0
755 find(PoolDescValues.begin(), PoolDescValues.end(), LoadAddr) !=
756 PoolDescValues.end()) {
758 assert("Make sure it's a load of the pool base, not a chaining field" &&
759 LI->getOperand(1) == Constant::getNullValue(Type::UIntTy) &&
760 LI->getOperand(2) == Constant::getNullValue(Type::UByteTy) &&
761 LI->getOperand(3) == Constant::getNullValue(Type::UByteTy));
763 // If it is a load of a pool base, keep track of it for future reference
764 PoolDescMap.insert(make_pair(LoadAddr, LI));
770 // If we run across a function call, forget all state... Calls to
771 // poolalloc/poolfree can invalidate the pool base pointer, so it should be
772 // reloaded the next time it is used. Furthermore, a call to a random
773 // function might call one of these functions, so be conservative. Through
774 // more analysis, this could be improved in the future.
776 void visitCallInst(CallInst *) {
781 static void addNodeMapping(DSNode *SrcNode, const PointerValSet &PVS,
782 map<DSNode*, PointerValSet> &NodeMapping) {
783 for (unsigned i = 0, e = PVS.size(); i != e; ++i)
784 if (NodeMapping[SrcNode].add(PVS[i])) { // Not in map yet?
785 assert(PVS[i].Index == 0 && "Node indexing not supported yet!");
786 DSNode *DestNode = PVS[i].Node;
788 // Loop over all of the outgoing links in the mapped graph
789 for (unsigned l = 0, le = DestNode->getNumOutgoingLinks(); l != le; ++l) {
790 PointerValSet &SrcSet = SrcNode->getOutgoingLink(l);
791 const PointerValSet &DestSet = DestNode->getOutgoingLink(l);
793 // Add all of the node mappings now!
794 for (unsigned si = 0, se = SrcSet.size(); si != se; ++si) {
795 assert(SrcSet[si].Index == 0 && "Can't handle node offset!");
796 addNodeMapping(SrcSet[si].Node, DestSet, NodeMapping);
802 // CalculateNodeMapping - There is a partial isomorphism between the graph
803 // passed in and the graph that is actually used by the function. We need to
804 // figure out what this mapping is so that we can transformFunctionBody the
805 // instructions in the function itself. Note that every node in the graph that
806 // we are interested in must be both in the local graph of the called function,
807 // and in the local graph of the calling function. Because of this, we only
808 // define the mapping for these nodes [conveniently these are the only nodes we
809 // CAN define a mapping for...]
811 // The roots of the graph that we are transforming is rooted in the arguments
812 // passed into the function from the caller. This is where we start our
813 // mapping calculation.
815 // The NodeMapping calculated maps from the callers graph to the called graph.
817 static void CalculateNodeMapping(Function *F, TransformFunctionInfo &TFI,
818 FunctionDSGraph &CallerGraph,
819 FunctionDSGraph &CalledGraph,
820 map<DSNode*, PointerValSet> &NodeMapping) {
822 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
823 // Figure out what nodes in the called graph the TFI.ArgInfo[i].Node node
826 // Only consider first node of sequence. Extra nodes may may be added
827 // to the TFI if the data structure requires more nodes than just the
828 // one the argument points to. We are only interested in the one the
829 // argument points to though.
831 if (TFI.ArgInfo[i].ArgNo != LastArgNo) {
832 if (TFI.ArgInfo[i].ArgNo == -1) {
833 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getRetNodes(),
836 // Figure out which node argument # ArgNo points to in the called graph.
837 Value *Arg = F->getArgumentList()[TFI.ArgInfo[i].ArgNo];
838 addNodeMapping(TFI.ArgInfo[i].Node, CalledGraph.getValueMap()[Arg],
841 LastArgNo = TFI.ArgInfo[i].ArgNo;
849 // addCallInfo - For a specified function call CI, figure out which pool
850 // descriptors need to be passed in as arguments, and which arguments need to be
851 // transformed into indices. If Arg != -1, the specified call argument is
852 // passed in as a pointer to a data structure.
854 void TransformFunctionInfo::addCallInfo(DataStructure *DS, CallInst *CI,
855 int Arg, DSNode *GraphNode,
856 map<DSNode*, PoolInfo> &PoolDescs) {
857 assert(CI->getCalledFunction() && "Cannot handle indirect calls yet!");
858 assert(Func == 0 || Func == CI->getCalledFunction() &&
859 "Function call record should always call the same function!");
860 assert(Call == 0 || Call == CI &&
861 "Call element already filled in with different value!");
862 Func = CI->getCalledFunction();
864 //FunctionDSGraph &CalledGraph = DS->getClosedDSGraph(Func);
866 // For now, add the entire graph that is pointed to by the call argument.
867 // This graph can and should be pruned to only what the function itself will
868 // use, because often this will be a dramatically smaller subset of what we
871 // FIXME: This should use pool links instead of extra arguments!
873 for (df_iterator<DSNode*> I = df_begin(GraphNode), E = df_end(GraphNode);
875 ArgInfo.push_back(CallArgInfo(Arg, *I, PoolDescs[*I].Handle));
878 static void markReachableNodes(const PointerValSet &Vals,
879 set<DSNode*> &ReachableNodes) {
880 for (unsigned n = 0, ne = Vals.size(); n != ne; ++n) {
881 DSNode *N = Vals[n].Node;
882 if (ReachableNodes.count(N) == 0) // Haven't already processed node?
883 ReachableNodes.insert(df_begin(N), df_end(N)); // Insert all
887 // Make sure that all dependant arguments are added to this transformation info.
888 // For example, if we call foo(null, P) and foo treats it's first and second
889 // arguments as belonging to the same data structure, the we MUST add entries to
890 // know that the null needs to be transformed into an index as well.
892 void TransformFunctionInfo::ensureDependantArgumentsIncluded(DataStructure *DS,
893 map<DSNode*, PoolInfo> &PoolDescs) {
894 // FIXME: This does not work for indirect function calls!!!
895 if (Func == 0) return; // FIXME!
897 // Make sure argument entries are sorted.
898 finalizeConstruction();
900 // Loop over the function signature, checking to see if there are any pointer
901 // arguments that we do not convert... if there is something we haven't
902 // converted, set done to false.
906 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
907 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
908 // We DO transform the ret val... skip all possible entries for retval
909 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
915 for (unsigned i = 0, e = Func->getArgumentList().size(); i != e; ++i) {
916 Argument *Arg = Func->getArgumentList()[i];
917 if (isa<PointerType>(Arg->getType())) {
918 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
919 // We DO transform this arg... skip all possible entries for argument
920 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
929 // If we already have entries for all pointer arguments and retvals, there
930 // certainly is no work to do. Bail out early to avoid building relatively
931 // expensive data structures.
935 #ifdef DEBUG_TRANSFORM_PROGRESS
936 cerr << "Must ensure dependant arguments for: " << Func->getName() << "\n";
939 // Otherwise, we MIGHT have to add the arguments/retval if they are part of
940 // the same datastructure graph as some other argument or retval that we ARE
943 // Get the data structure graph for the called function.
945 FunctionDSGraph &CalledDS = DS->getClosedDSGraph(Func);
947 // Build a mapping between the nodes in our current graph and the nodes in the
948 // called function's graph. We build it based on our _incomplete_
949 // transformation information, because it contains all of the info that we
952 map<DSNode*, PointerValSet> NodeMapping;
953 CalculateNodeMapping(Func, *this,
954 DS->getClosedDSGraph(Call->getParent()->getParent()),
955 CalledDS, NodeMapping);
957 // Build the inverted version of the node mapping, that maps from a node in
958 // the called functions graph to a single node in the caller graph.
960 map<DSNode*, DSNode*> InverseNodeMap;
961 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin(),
962 E = NodeMapping.end(); I != E; ++I) {
963 PointerValSet &CalledNodes = I->second;
964 for (unsigned i = 0, e = CalledNodes.size(); i != e; ++i)
965 InverseNodeMap[CalledNodes[i].Node] = I->first;
967 NodeMapping.clear(); // Done with information, free memory
969 // Build a set of reachable nodes from the arguments/retval that we ARE
971 set<DSNode*> ReachableNodes;
973 // Loop through all of the arguments, marking all of the reachable data
974 // structure nodes reachable if they are from this pointer...
976 for (unsigned i = 0, e = ArgInfo.size(); i != e; ++i) {
977 if (ArgInfo[i].ArgNo == -1) {
978 if (i == 0) // Only process retvals once (performance opt)
979 markReachableNodes(CalledDS.getRetNodes(), ReachableNodes);
980 } else { // If it's an argument value...
981 Argument *Arg = Func->getArgumentList()[ArgInfo[i].ArgNo];
982 if (isa<PointerType>(Arg->getType()))
983 markReachableNodes(CalledDS.getValueMap()[Arg], ReachableNodes);
987 // Now that we know which nodes are already reachable, see if any of the
988 // arguments that we are not passing values in for can reach one of the
992 // <FIXME> IN THEORY, we should allow arbitrary paths from the argument to
993 // nodes we know about. The problem is that if we do this, then I don't know
994 // how to get pool pointers for this head list. Since we are completely
995 // deadline driven, I'll just allow direct accesses to the graph. </FIXME>
999 if (isa<PointerType>(Func->getReturnType())) // Make sure we convert retval
1000 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == -1) {
1001 // We DO transform the ret val... skip all possible entries for retval
1002 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == -1)
1005 // See what the return value points to...
1007 // FIXME: This should generalize to any number of nodes, just see if any
1009 assert(CalledDS.getRetNodes().size() == 1 &&
1010 "Assumes only one node is returned");
1011 DSNode *N = CalledDS.getRetNodes()[0].Node;
1013 // If the return value is not marked as being passed in, but it NEEDS to
1014 // be transformed, then make it known now.
1016 if (ReachableNodes.count(N)) {
1017 #ifdef DEBUG_TRANSFORM_PROGRESS
1018 cerr << "ensure dependant arguments adds return value entry!\n";
1020 addCallInfo(DS, Call, -1, InverseNodeMap[N], PoolDescs);
1023 finalizeConstruction();
1027 for (unsigned i = 0, e = Func->getArgumentList().size(); i != e; ++i) {
1028 Argument *Arg = Func->getArgumentList()[i];
1029 if (isa<PointerType>(Arg->getType())) {
1030 if (PtrNo < ArgInfo.size() && ArgInfo[PtrNo++].ArgNo == (int)i) {
1031 // We DO transform this arg... skip all possible entries for argument
1032 while (PtrNo < ArgInfo.size() && ArgInfo[PtrNo].ArgNo == (int)i)
1035 // This should generalize to any number of nodes, just see if any are
1037 assert(CalledDS.getValueMap()[Arg].size() == 1 &&
1038 "Only handle case where pointing to one node so far!");
1040 // If the arg is not marked as being passed in, but it NEEDS to
1041 // be transformed, then make it known now.
1043 DSNode *N = CalledDS.getValueMap()[Arg][0].Node;
1044 if (ReachableNodes.count(N)) {
1045 #ifdef DEBUG_TRANSFORM_PROGRESS
1046 cerr << "ensure dependant arguments adds for arg #" << i << "\n";
1048 addCallInfo(DS, Call, i, InverseNodeMap[N], PoolDescs);
1051 finalizeConstruction();
1059 // transformFunctionBody - This transforms the instruction in 'F' to use the
1060 // pools specified in PoolDescs when modifying data structure nodes specified in
1061 // the PoolDescs map. Specifically, scalar values specified in the Scalars
1062 // vector must be remapped. IPFGraph is the closed data structure graph for F,
1063 // of which the PoolDescriptor nodes come from.
1065 void PoolAllocate::transformFunctionBody(Function *F, FunctionDSGraph &IPFGraph,
1066 map<DSNode*, PoolInfo> &PoolDescs) {
1068 // Loop through the value map looking for scalars that refer to nonescaping
1069 // allocations. Add them to the Scalars vector. Note that we may have
1070 // multiple entries in the Scalars vector for each value if it points to more
1073 map<Value*, PointerValSet> &ValMap = IPFGraph.getValueMap();
1074 vector<ScalarInfo> Scalars;
1076 #ifdef DEBUG_TRANSFORM_PROGRESS
1077 cerr << "Building scalar map for fn '" << F->getName() << "' body:\n";
1080 for (map<Value*, PointerValSet>::iterator I = ValMap.begin(),
1081 E = ValMap.end(); I != E; ++I) {
1082 const PointerValSet &PVS = I->second; // Set of things pointed to by scalar
1084 // Check to see if the scalar points to a data structure node...
1085 for (unsigned i = 0, e = PVS.size(); i != e; ++i) {
1086 if (PVS[i].Index) { cerr << "Problem in " << F->getName() << " for " << I->first << "\n"; }
1087 assert(PVS[i].Index == 0 && "Nonzero not handled yet!");
1089 // If the allocation is in the nonescaping set...
1090 map<DSNode*, PoolInfo>::iterator AI = PoolDescs.find(PVS[i].Node);
1091 if (AI != PoolDescs.end()) { // Add it to the list of scalars
1092 Scalars.push_back(ScalarInfo(I->first, AI->second));
1093 #ifdef DEBUG_TRANSFORM_PROGRESS
1094 cerr << "\nScalar Mapping from:" << I->first
1095 << "Scalar Mapping to: "; PVS.print(cerr);
1101 #ifdef DEBUG_TRANSFORM_PROGRESS
1102 cerr << "\nIn '" << F->getName()
1103 << "': Found the following values that point to poolable nodes:\n";
1105 for (unsigned i = 0, e = Scalars.size(); i != e; ++i)
1106 cerr << Scalars[i].Val;
1110 // CallMap - Contain an entry for every call instruction that needs to be
1111 // transformed. Each entry in the map contains information about what we need
1112 // to do to each call site to change it to work.
1114 map<CallInst*, TransformFunctionInfo> CallMap;
1116 // Now we need to figure out what called functions we need to transform, and
1117 // how. To do this, we look at all of the scalars, seeing which functions are
1118 // either used as a scalar value (so they return a data structure), or are
1119 // passed one of our scalar values.
1121 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1122 Value *ScalarVal = Scalars[i].Val;
1124 // Check to see if the scalar _IS_ a call...
1125 if (CallInst *CI = dyn_cast<CallInst>(ScalarVal))
1126 // If so, add information about the pool it will be returning...
1127 CallMap[CI].addCallInfo(DS, CI, -1, Scalars[i].Pool.Node, PoolDescs);
1129 // Check to see if the scalar is an operand to a call...
1130 for (Value::use_iterator UI = ScalarVal->use_begin(),
1131 UE = ScalarVal->use_end(); UI != UE; ++UI) {
1132 if (CallInst *CI = dyn_cast<CallInst>(*UI)) {
1133 // Find out which operand this is to the call instruction...
1134 User::op_iterator OI = find(CI->op_begin(), CI->op_end(), ScalarVal);
1135 assert(OI != CI->op_end() && "Call on use list but not an operand!?");
1136 assert(OI != CI->op_begin() && "Pointer operand is call destination?");
1138 // FIXME: This is broken if the same pointer is passed to a call more
1139 // than once! It will get multiple entries for the first pointer.
1141 // Add the operand number and pool handle to the call table...
1142 CallMap[CI].addCallInfo(DS, CI, OI-CI->op_begin()-1,
1143 Scalars[i].Pool.Node, PoolDescs);
1148 // Make sure that all dependant arguments are added as well. For example, if
1149 // we call foo(null, P) and foo treats it's first and second arguments as
1150 // belonging to the same data structure, the we MUST set up the CallMap to
1151 // know that the null needs to be transformed into an index as well.
1153 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1154 I != CallMap.end(); ++I)
1155 I->second.ensureDependantArgumentsIncluded(DS, PoolDescs);
1157 #ifdef DEBUG_TRANSFORM_PROGRESS
1158 // Print out call map...
1159 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin();
1160 I != CallMap.end(); ++I) {
1161 cerr << "For call: " << I->first;
1162 cerr << I->second.Func->getName() << " must pass pool pointer for args #";
1163 for (unsigned i = 0; i < I->second.ArgInfo.size(); ++i)
1164 cerr << I->second.ArgInfo[i].ArgNo << ", ";
1169 // Loop through all of the call nodes, recursively creating the new functions
1170 // that we want to call... This uses a map to prevent infinite recursion and
1171 // to avoid duplicating functions unneccesarily.
1173 for (map<CallInst*, TransformFunctionInfo>::iterator I = CallMap.begin(),
1174 E = CallMap.end(); I != E; ++I) {
1175 // Transform all of the functions we need, or at least ensure there is a
1176 // cached version available.
1177 transformFunction(I->second, IPFGraph, PoolDescs);
1180 // Now that all of the functions that we want to call are available, transform
1181 // the local function so that it uses the pools locally and passes them to the
1182 // functions that we just hacked up.
1185 // First step, find the instructions to be modified.
1186 vector<Instruction*> InstToFix;
1187 for (unsigned i = 0, e = Scalars.size(); i != e; ++i) {
1188 Value *ScalarVal = Scalars[i].Val;
1190 // Check to see if the scalar _IS_ an instruction. If so, it is involved.
1191 if (Instruction *Inst = dyn_cast<Instruction>(ScalarVal))
1192 InstToFix.push_back(Inst);
1194 // All all of the instructions that use the scalar as an operand...
1195 for (Value::use_iterator UI = ScalarVal->use_begin(),
1196 UE = ScalarVal->use_end(); UI != UE; ++UI)
1197 InstToFix.push_back(cast<Instruction>(*UI));
1200 // Make sure that we get return instructions that return a null value from the
1203 if (!IPFGraph.getRetNodes().empty()) {
1204 assert(IPFGraph.getRetNodes().size() == 1 && "Can only return one node?");
1205 PointerVal RetNode = IPFGraph.getRetNodes()[0];
1206 assert(RetNode.Index == 0 && "Subindexing not implemented yet!");
1208 // Only process return instructions if the return value of this function is
1209 // part of one of the data structures we are transforming...
1211 if (PoolDescs.count(RetNode.Node)) {
1212 // Loop over all of the basic blocks, adding return instructions...
1213 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1214 if (ReturnInst *RI = dyn_cast<ReturnInst>((*I)->getTerminator()))
1215 InstToFix.push_back(RI);
1221 // Eliminate duplicates by sorting, then removing equal neighbors.
1222 sort(InstToFix.begin(), InstToFix.end());
1223 InstToFix.erase(unique(InstToFix.begin(), InstToFix.end()), InstToFix.end());
1225 // Loop over all of the instructions to transform, creating the new
1226 // replacement instructions for them. This also unlinks them from the
1227 // function so they can be safely deleted later.
1229 map<Value*, Value*> XFormMap;
1230 NewInstructionCreator NIC(*this, Scalars, CallMap, XFormMap);
1232 // Visit all instructions... creating the new instructions that we need and
1233 // unlinking the old instructions from the function...
1235 #ifdef DEBUG_TRANSFORM_PROGRESS
1236 for (unsigned i = 0, e = InstToFix.size(); i != e; ++i) {
1237 cerr << "Fixing: " << InstToFix[i];
1238 NIC.visit(InstToFix[i]);
1241 NIC.visit(InstToFix.begin(), InstToFix.end());
1244 // Make all instructions we will delete "let go" of their operands... so that
1245 // we can safely delete Arguments whose types have changed...
1247 for_each(InstToFix.begin(), InstToFix.end(),
1248 mem_fun(&Instruction::dropAllReferences));
1250 // Loop through all of the pointer arguments coming into the function,
1251 // replacing them with arguments of POINTERTYPE to match the function type of
1254 FunctionType::ParamTypes::const_iterator TI =
1255 F->getFunctionType()->getParamTypes().begin();
1256 for (Function::ArgumentListType::iterator I = F->getArgumentList().begin(),
1257 E = F->getArgumentList().end(); I != E; ++I, ++TI) {
1259 if (Arg->getType() != *TI) {
1260 assert(isa<PointerType>(Arg->getType()) && *TI == POINTERTYPE);
1261 Argument *NewArg = new Argument(*TI, Arg->getName());
1262 XFormMap[Arg] = NewArg; // Map old arg into new arg...
1264 // Replace the old argument and then delete it...
1265 delete F->getArgumentList().replaceWith(I, NewArg);
1269 // Now that all of the new instructions have been created, we can update all
1270 // of the references to dummy values to be references to the actual values
1271 // that are computed.
1273 NIC.updateReferences();
1275 #ifdef DEBUG_TRANSFORM_PROGRESS
1276 cerr << "TRANSFORMED FUNCTION:\n" << F;
1279 // Delete all of the "instructions to fix"
1280 for_each(InstToFix.begin(), InstToFix.end(), deleter<Instruction>);
1282 // Eliminate pool base loads that we can easily prove are redundant
1284 PoolBaseLoadEliminator(PoolDescs).visit(F);
1286 // Since we have liberally hacked the function to pieces, we want to inform
1287 // the datastructure pass that its internal representation is out of date.
1289 DS->invalidateFunction(F);
1294 // transformFunction - Transform the specified function the specified way. It
1295 // we have already transformed that function that way, don't do anything. The
1296 // nodes in the TransformFunctionInfo come out of callers data structure graph.
1298 void PoolAllocate::transformFunction(TransformFunctionInfo &TFI,
1299 FunctionDSGraph &CallerIPGraph,
1300 map<DSNode*, PoolInfo> &CallerPoolDesc) {
1301 if (getTransformedFunction(TFI)) return; // Function xformation already done?
1303 #ifdef DEBUG_TRANSFORM_PROGRESS
1304 cerr << "********** Entering transformFunction for "
1305 << TFI.Func->getName() << ":\n";
1306 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i)
1307 cerr << " ArgInfo[" << i << "] = " << TFI.ArgInfo[i].ArgNo << "\n";
1311 const FunctionType *OldFuncType = TFI.Func->getFunctionType();
1313 assert(!OldFuncType->isVarArg() && "Vararg functions not handled yet!");
1315 // Build the type for the new function that we are transforming
1316 vector<const Type*> ArgTys;
1317 ArgTys.reserve(OldFuncType->getNumParams()+TFI.ArgInfo.size());
1318 for (unsigned i = 0, e = OldFuncType->getNumParams(); i != e; ++i)
1319 ArgTys.push_back(OldFuncType->getParamType(i));
1321 const Type *RetType = OldFuncType->getReturnType();
1323 // Add one pool pointer for every argument that needs to be supplemented.
1324 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1325 if (TFI.ArgInfo[i].ArgNo == -1)
1326 RetType = POINTERTYPE; // Return a pointer
1328 ArgTys[TFI.ArgInfo[i].ArgNo] = POINTERTYPE; // Pass a pointer
1329 ArgTys.push_back(PointerType::get(CallerPoolDesc.find(TFI.ArgInfo[i].Node)
1330 ->second.PoolType));
1333 // Build the new function type...
1334 const FunctionType *NewFuncType = FunctionType::get(RetType, ArgTys,
1335 OldFuncType->isVarArg());
1337 // The new function is internal, because we know that only we can call it.
1338 // This also helps subsequent IP transformations to eliminate duplicated pool
1339 // pointers (which look like the same value is always passed into a parameter,
1340 // allowing it to be easily eliminated).
1342 Function *NewFunc = new Function(NewFuncType, true,
1343 TFI.Func->getName()+".poolxform");
1344 CurModule->getFunctionList().push_back(NewFunc);
1347 #ifdef DEBUG_TRANSFORM_PROGRESS
1348 cerr << "Created function prototype: " << NewFunc << "\n";
1351 // Add the newly formed function to the TransformedFunctions table so that
1352 // infinite recursion does not occur!
1354 TransformedFunctions[TFI] = NewFunc;
1356 // Add arguments to the function... starting with all of the old arguments
1357 vector<Value*> ArgMap;
1358 for (unsigned i = 0, e = TFI.Func->getArgumentList().size(); i != e; ++i) {
1359 const Argument *OFA = TFI.Func->getArgumentList()[i];
1360 Argument *NFA = new Argument(OFA->getType(), OFA->getName());
1361 NewFunc->getArgumentList().push_back(NFA);
1362 ArgMap.push_back(NFA); // Keep track of the arguments
1365 // Now add all of the arguments corresponding to pools passed in...
1366 for (unsigned i = 0, e = TFI.ArgInfo.size(); i != e; ++i) {
1367 CallArgInfo &AI = TFI.ArgInfo[i];
1372 Name = ArgMap[AI.ArgNo]->getName(); // Get the arg name
1373 const Type *Ty = PointerType::get(CallerPoolDesc[AI.Node].PoolType);
1374 Argument *NFA = new Argument(Ty, Name+".pool");
1375 NewFunc->getArgumentList().push_back(NFA);
1378 // Now clone the body of the old function into the new function...
1379 CloneFunctionInto(NewFunc, TFI.Func, ArgMap);
1381 // Okay, now we have a function that is identical to the old one, except that
1382 // it has extra arguments for the pools coming in. Now we have to get the
1383 // data structure graph for the function we are replacing, and figure out how
1384 // our graph nodes map to the graph nodes in the dest function.
1386 FunctionDSGraph &DSGraph = DS->getClosedDSGraph(NewFunc);
1388 // NodeMapping - Multimap from callers graph to called graph. We are
1389 // guaranteed that the called function graph has more nodes than the caller,
1390 // or exactly the same number of nodes. This is because the called function
1391 // might not know that two nodes are merged when considering the callers
1392 // context, but the caller obviously does. Because of this, a single node in
1393 // the calling function's data structure graph can map to multiple nodes in
1394 // the called functions graph.
1396 map<DSNode*, PointerValSet> NodeMapping;
1398 CalculateNodeMapping(NewFunc, TFI, CallerIPGraph, DSGraph,
1401 // Print out the node mapping...
1402 #ifdef DEBUG_TRANSFORM_PROGRESS
1403 cerr << "\nNode mapping for call of " << NewFunc->getName() << "\n";
1404 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1405 I != NodeMapping.end(); ++I) {
1406 cerr << "Map: "; I->first->print(cerr);
1407 cerr << "To: "; I->second.print(cerr);
1412 // Fill in the PoolDescriptor information for the transformed function so that
1413 // it can determine which value holds the pool descriptor for each data
1414 // structure node that it accesses.
1416 map<DSNode*, PoolInfo> PoolDescs;
1418 #ifdef DEBUG_TRANSFORM_PROGRESS
1419 cerr << "\nCalculating the pool descriptor map:\n";
1422 // Calculate as much of the pool descriptor map as possible. Since we have
1423 // the node mapping between the caller and callee functions, and we have the
1424 // pool descriptor information of the caller, we can calculate a partical pool
1425 // descriptor map for the called function.
1427 // The nodes that we do not have complete information for are the ones that
1428 // are accessed by loading pointers derived from arguments passed in, but that
1429 // are not passed in directly. In this case, we have all of the information
1430 // except a pool value. If the called function refers to this pool, the pool
1431 // value will be loaded from the pool graph and added to the map as neccesary.
1433 for (map<DSNode*, PointerValSet>::iterator I = NodeMapping.begin();
1434 I != NodeMapping.end(); ++I) {
1435 DSNode *CallerNode = I->first;
1436 PoolInfo &CallerPI = CallerPoolDesc[CallerNode];
1438 // Check to see if we have a node pointer passed in for this value...
1439 Value *CalleeValue = 0;
1440 for (unsigned a = 0, ae = TFI.ArgInfo.size(); a != ae; ++a)
1441 if (TFI.ArgInfo[a].Node == CallerNode) {
1442 // Calculate the argument number that the pool is to the function
1443 // call... The call instruction should not have the pool operands added
1445 unsigned ArgNo = TFI.Call->getNumOperands()-1+a;
1446 #ifdef DEBUG_TRANSFORM_PROGRESS
1447 cerr << "Should be argument #: " << ArgNo << "[i = " << a << "]\n";
1449 assert(ArgNo < NewFunc->getArgumentList().size() &&
1450 "Call already has pool arguments added??");
1452 // Map the pool argument into the called function...
1453 CalleeValue = NewFunc->getArgumentList()[ArgNo];
1454 break; // Found value, quit loop
1457 // Loop over all of the data structure nodes that this incoming node maps to
1458 // Creating a PoolInfo structure for them.
1459 for (unsigned i = 0, e = I->second.size(); i != e; ++i) {
1460 assert(I->second[i].Index == 0 && "Doesn't handle subindexing yet!");
1461 DSNode *CalleeNode = I->second[i].Node;
1463 // Add the descriptor. We already know everything about it by now, much
1464 // of it is the same as the caller info.
1466 PoolDescs.insert(make_pair(CalleeNode,
1467 PoolInfo(CalleeNode, CalleeValue,
1469 CallerPI.PoolType)));
1473 // We must destroy the node mapping so that we don't have latent references
1474 // into the data structure graph for the new function. Otherwise we get
1475 // assertion failures when transformFunctionBody tries to invalidate the
1478 NodeMapping.clear();
1480 // Now that we know everything we need about the function, transform the body
1483 transformFunctionBody(NewFunc, DSGraph, PoolDescs);
1485 #ifdef DEBUG_TRANSFORM_PROGRESS
1486 cerr << "Function after transformation:\n" << NewFunc;
1490 static unsigned countPointerTypes(const Type *Ty) {
1491 if (isa<PointerType>(Ty)) {
1493 } else if (StructType *STy = dyn_cast<StructType>(Ty)) {
1495 for (unsigned i = 0, e = STy->getElementTypes().size(); i != e; ++i)
1496 Num += countPointerTypes(STy->getElementTypes()[i]);
1498 } else if (ArrayType *ATy = dyn_cast<ArrayType>(Ty)) {
1499 return countPointerTypes(ATy->getElementType());
1501 assert(Ty->isPrimitiveType() && "Unknown derived type!");
1506 // CreatePools - Insert instructions into the function we are processing to
1507 // create all of the memory pool objects themselves. This also inserts
1508 // destruction code. Add an alloca for each pool that is allocated to the
1509 // PoolDescs vector.
1511 void PoolAllocate::CreatePools(Function *F, const vector<AllocDSNode*> &Allocs,
1512 map<DSNode*, PoolInfo> &PoolDescs) {
1513 // Find all of the return nodes in the function...
1514 vector<BasicBlock*> ReturnNodes;
1515 for (Function::iterator I = F->begin(), E = F->end(); I != E; ++I)
1516 if (isa<ReturnInst>((*I)->getTerminator()))
1517 ReturnNodes.push_back(*I);
1519 #ifdef DEBUG_CREATE_POOLS
1520 cerr << "Allocs that we are pool allocating:\n";
1521 for (unsigned i = 0, e = Allocs.size(); i != e; ++i)
1525 map<DSNode*, PATypeHolder> AbsPoolTyMap;
1527 // First pass over the allocations to process...
1528 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1529 // Create the pooldescriptor mapping... with null entries for everything
1530 // except the node & NewType fields.
1532 map<DSNode*, PoolInfo>::iterator PI =
1533 PoolDescs.insert(make_pair(Allocs[i], PoolInfo(Allocs[i]))).first;
1535 // Add a symbol table entry for the new type if there was one for the old
1537 string OldName = CurModule->getTypeName(Allocs[i]->getType());
1538 if (OldName.empty()) OldName = "node";
1539 CurModule->addTypeName(OldName+".p", PI->second.NewType);
1541 // Create the abstract pool types that will need to be resolved in a second
1542 // pass once an abstract type is created for each pool.
1544 // Can only handle limited shapes for now...
1545 const Type *OldNodeTy = Allocs[i]->getType();
1546 vector<const Type*> PoolTypes;
1548 // Pool type is the first element of the pool descriptor type...
1549 PoolTypes.push_back(getPoolType(PoolDescs[Allocs[i]].NewType));
1551 unsigned NumPointers = countPointerTypes(OldNodeTy);
1552 while (NumPointers--) // Add a different opaque type for each pointer
1553 PoolTypes.push_back(OpaqueType::get());
1555 assert(Allocs[i]->getNumLinks() == PoolTypes.size()-1 &&
1556 "Node should have same number of pointers as pool!");
1558 StructType *PoolType = StructType::get(PoolTypes);
1560 // Add a symbol table entry for the pooltype if possible...
1561 CurModule->addTypeName(OldName+".pool", PoolType);
1563 // Create the pool type, with opaque values for pointers...
1564 AbsPoolTyMap.insert(make_pair(Allocs[i], PoolType));
1565 #ifdef DEBUG_CREATE_POOLS
1566 cerr << "POOL TY: " << AbsPoolTyMap.find(Allocs[i])->second.get() << "\n";
1570 // Now that we have types for all of the pool types, link them all together.
1571 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1572 PATypeHolder &PoolTyH = AbsPoolTyMap.find(Allocs[i])->second;
1574 // Resolve all of the outgoing pointer types of this pool node...
1575 for (unsigned p = 0, pe = Allocs[i]->getNumLinks(); p != pe; ++p) {
1576 PointerValSet &PVS = Allocs[i]->getLink(p);
1577 assert(!PVS.empty() && "Outgoing edge is empty, field unused, can"
1578 " probably just leave the type opaque or something dumb.");
1580 for (Out = 0; AbsPoolTyMap.count(PVS[Out].Node) == 0; ++Out)
1581 assert(Out != PVS.size() && "No edge to an outgoing allocation node!?");
1583 assert(PVS[Out].Index == 0 && "Subindexing not implemented yet!");
1585 // The actual struct type could change each time through the loop, so it's
1586 // NOT loop invariant.
1587 StructType *PoolTy = cast<StructType>(PoolTyH.get());
1589 // Get the opaque type...
1591 cast<DerivedType>(PoolTy->getElementTypes()[p+1].get());
1593 #ifdef DEBUG_CREATE_POOLS
1594 cerr << "Refining " << ElTy << " of " << PoolTy << " to "
1595 << AbsPoolTyMap.find(PVS[Out].Node)->second.get() << "\n";
1598 const Type *RefPoolTy = AbsPoolTyMap.find(PVS[Out].Node)->second.get();
1599 ElTy->refineAbstractTypeTo(PointerType::get(RefPoolTy));
1601 #ifdef DEBUG_CREATE_POOLS
1602 cerr << "Result pool type is: " << PoolTyH.get() << "\n";
1607 // Create the code that goes in the entry and exit nodes for the function...
1608 vector<Instruction*> EntryNodeInsts;
1609 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1610 PoolInfo &PI = PoolDescs[Allocs[i]];
1612 // Fill in the pool type for this pool...
1613 PI.PoolType = AbsPoolTyMap.find(Allocs[i])->second.get();
1614 assert(!PI.PoolType->isAbstract() &&
1615 "Pool type should not be abstract anymore!");
1617 // Add an allocation and a free for each pool...
1618 AllocaInst *PoolAlloc
1619 = new AllocaInst(PointerType::get(PI.PoolType), 0,
1620 CurModule->getTypeName(PI.PoolType));
1621 PI.Handle = PoolAlloc;
1622 EntryNodeInsts.push_back(PoolAlloc);
1623 AllocationInst *AI = Allocs[i]->getAllocation();
1625 // Initialize the pool. We need to know how big each allocation is. For
1626 // our purposes here, we assume we are allocating a scalar, or array of
1629 unsigned ElSize = TargetData.getTypeSize(PI.NewType);
1631 vector<Value*> Args;
1632 Args.push_back(ConstantUInt::get(Type::UIntTy, ElSize));
1633 Args.push_back(PoolAlloc); // Pool to initialize
1634 EntryNodeInsts.push_back(new CallInst(PoolInit, Args));
1636 // Add code to destroy the pool in all of the exit nodes of the function...
1638 Args.push_back(PoolAlloc); // Pool to initialize
1640 for (unsigned EN = 0, ENE = ReturnNodes.size(); EN != ENE; ++EN) {
1641 Instruction *Destroy = new CallInst(PoolDestroy, Args);
1643 // Insert it before the return instruction...
1644 BasicBlock *RetNode = ReturnNodes[EN];
1645 RetNode->getInstList().insert(RetNode->end()-1, Destroy);
1649 // Now that all of the pool descriptors have been created, link them together
1650 // so that called functions can get links as neccesary...
1652 for (unsigned i = 0, e = Allocs.size(); i != e; ++i) {
1653 PoolInfo &PI = PoolDescs[Allocs[i]];
1655 // For every pointer in the data structure, initialize a link that
1656 // indicates which pool to access...
1658 vector<Value*> Indices(2);
1659 Indices[0] = ConstantUInt::get(Type::UIntTy, 0);
1660 for (unsigned l = 0, le = PI.Node->getNumLinks(); l != le; ++l)
1661 // Only store an entry for the field if the field is used!
1662 if (!PI.Node->getLink(l).empty()) {
1663 assert(PI.Node->getLink(l).size() == 1 && "Should have only one link!");
1664 PointerVal PV = PI.Node->getLink(l)[0];
1665 assert(PV.Index == 0 && "Subindexing not supported yet!");
1666 PoolInfo &LinkedPool = PoolDescs[PV.Node];
1667 Indices[1] = ConstantUInt::get(Type::UByteTy, 1+l);
1669 EntryNodeInsts.push_back(new StoreInst(LinkedPool.Handle, PI.Handle,
1674 // Insert the entry node code into the entry block...
1675 F->getEntryNode()->getInstList().insert(F->getEntryNode()->begin()+1,
1676 EntryNodeInsts.begin(),
1677 EntryNodeInsts.end());
1681 // addPoolPrototypes - Add prototypes for the pool functions to the specified
1682 // module and update the Pool* instance variables to point to them.
1684 void PoolAllocate::addPoolPrototypes(Module *M) {
1685 // Get poolinit function...
1686 vector<const Type*> Args;
1687 Args.push_back(Type::UIntTy); // Num bytes per element
1688 FunctionType *PoolInitTy = FunctionType::get(Type::VoidTy, Args, true);
1689 PoolInit = M->getOrInsertFunction("poolinit", PoolInitTy);
1691 // Get pooldestroy function...
1692 Args.pop_back(); // Only takes a pool...
1693 FunctionType *PoolDestroyTy = FunctionType::get(Type::VoidTy, Args, true);
1694 PoolDestroy = M->getOrInsertFunction("pooldestroy", PoolDestroyTy);
1696 // Get the poolalloc function...
1697 FunctionType *PoolAllocTy = FunctionType::get(POINTERTYPE, Args, true);
1698 PoolAlloc = M->getOrInsertFunction("poolalloc", PoolAllocTy);
1700 // Get the poolfree function...
1701 Args.push_back(POINTERTYPE); // Pointer to free
1702 FunctionType *PoolFreeTy = FunctionType::get(Type::VoidTy, Args, true);
1703 PoolFree = M->getOrInsertFunction("poolfree", PoolFreeTy);
1705 Args[0] = Type::UIntTy; // Number of slots to allocate
1706 FunctionType *PoolAllocArrayTy = FunctionType::get(POINTERTYPE, Args, true);
1707 PoolAllocArray = M->getOrInsertFunction("poolallocarray", PoolAllocArrayTy);
1711 bool PoolAllocate::run(Module *M) {
1712 addPoolPrototypes(M);
1715 DS = &getAnalysis<DataStructure>();
1716 bool Changed = false;
1718 // We cannot use an iterator here because it will get invalidated when we add
1719 // functions to the module later...
1720 for (unsigned i = 0; i != M->size(); ++i)
1721 if (!M->getFunctionList()[i]->isExternal()) {
1722 Changed |= processFunction(M->getFunctionList()[i]);
1724 cerr << "Only processing one function\n";
1735 // createPoolAllocatePass - Global function to access the functionality of this
1738 Pass *createPoolAllocatePass() { return new PoolAllocate(); }